ROBOCREW

Expectation vs. Reality

The companies that will win this aren't the ones making the flashiest demo videos. They're the ones locking down motor supply chains, shipping wheeled-arm units to factories TODAY, and building the software ecosystem on top. The humanoid robot industry just shifted from "will it work" to "who ships first at scale." And most of the world hasn't noticed yet.

And at the component level, a quiet revolution is happening. GaN power FETs are replacing silicon MOSFETs inside robot joint actuators. This means: → 50% smaller power stages → Switching frequencies jumping from kHz to MHz → Dramatically smoother motor control GaN market for robotics: $260M (2023) → $2B (2029). CAGR 40.6%.

Everyone's debating whether humanoid robots will work. Meanwhile in China, Q1 2025 orders for most humanoid companies already exceeded their entire 2024 year. Here's what I'm seeing on the ground that Western observers are completely missing 🧵

The business model is two-layered: → Generic atomic skills (open doors, grasp, deliver, navigate) get open-sourced to build ecosystem → Vertical data packages (logistics depalletizing, automotive assembly, SMT handling) get sold to enterprises for real money Three types of buyers are converging: manufacturers who need robots working NOW in their factories, foundation model companies hungry for multi-modal training data, and universities running benchmarks. The companies building real-robot data moats right now will define who wins the humanoid race. Not the ones making viral clips. Watch the data. That's where the real power is.

Here's the problem: all open-source humanoid datasets globally are still in the tens-of-thousands of hours. For large-model training, that's basically zero. That's why insiders are calling 2025 "Year Zero" for robotics data scaling. What most people are missing: China now has 14 dedicated humanoid robot training facilities producing 25M real-robot data entries/year 60,000+ minutes of data have been open-sourced, broken into 117 reusable atomic skills 20,000 hours of commercial data already delivered to enterprise buyers A national-level open-source data community just launched — backed by the state, not any single company Real-robot data is being issued IP certificates and asset insurance. It's moving from gray-area byproduct to financeable asset on balance sheets

Everyone's obsessing over humanoid robot hardware. Walking, dancing, backflips — all solved. But here's what I'm seeing on the ground in China that nobody outside is talking about: the real bottleneck is data. And the race to own it is quietly reshaping the entire industry.

The winners in this space won't be the ones with the best tricks on camera. They'll be the ones who make the unit economics work across real verticals. Follow along — we're tracking the robotics industry from both sides of the Pacific. 🤖

🔌 More Platform-Oriented The smartest companies are building a universal chassis + toolchain and letting integrators build vertical solutions on top. Power companies build inspection packages. Security firms build patrol networks. Think smartphones going from brands building their own apps → open ecosystems. One company already producing 30,000 joint modules/month with sub-2% failure rate.

The viral robot dog videos are impressive — but they don't tell the real story. a deep-dive breaking down where quadruped robots are actually headed in the next 1–2 years. The industry is shifting from flashy demos to real deployment. 4 key directions: → More Industrial — purpose-built tools for power grid inspection, security, firefighting → More All-Terrain — stairs, rubble, extreme weather, not just flat surfaces → More Intelligent — from RL-based locomotion to natural language task assignment → More Platform-Oriented — one chassis, infinite use cases (think: smartphone app stores) The companies that win won't be the ones making the coolest tricks. They'll be the ones where the unit economics work. Swipe for the full breakdown ↗️

MakerMods Hackathon Day 1 Recap

Researchers trained a humanoid robot to play tennis using only 5 hours of motion capture data The robot can now sustain multi-shot rallies with human players, hitting balls traveling >15 m/s with a ~90% success rate AlphaGo for every sport is coming

We built a cognitive application for humanoid robots. Not just what they're doing — why. PRISM shows you the full AI automated reasoning chain in real time: every decision, every rejected branch, every confidence score. Before the robot moves, you see the thought.